api-inference-community icon indicating copy to clipboard operation
api-inference-community copied to clipboard

update fairseq to stable commit

Open bayartsogt-ya opened this issue 3 years ago • 6 comments

Issue https://github.com/huggingface/api-inference-community/issues/134

Current version of fairseq is causing issues in STT pipeline so I changed it to the last commit of stable release of fairseq.

current: d47119871c2ac9a0a0aa2904dd8cfc1929b113d9 last: 59d966a92aabc68b6e0fe1f7bc3eeccbbbe91413

Error reproduce: https://www.kaggle.com/code/bayartsogtya/notebooke77335cbe2/notebook version 1 -> d471198 version 2 -> 59d966

link to huggingface failed example

bayartsogt-ya avatar Nov 03 '22 16:11 bayartsogt-ya

@osanseviero I wonder if there is any way to trigger automatic test in action from this fork?

bayartsogt-ya avatar Nov 03 '22 16:11 bayartsogt-ya

@bayartsogt-ya , the test is run. It's the red one.

The current problem is that the docker takes too long to spawn for the test to pass (it's because the docker downloads a lot of files even on subsequent reloads).

If you want to test manually if it works for model facebook/mymodel you can do

./manage.py docker facebook/mymodel and send data through curl to see the output.

Does that help ?

Narsil avatar Nov 03 '22 17:11 Narsil

Here is a log of fairseq-docker build action:

Running script /app/prestart.sh
2022-11-03 16:43:55 | INFO | fairseq.tasks.text_to_speech | Please install tensorboardX: pip install tensorboardX
Traceback (most recent call last):
  File "app/main.py", line 7, in <module>
    from app.pipelines import Pipeline, SpeechToSpeechPipeline, TextToSpeechPipeline
  File "/app/app/pipelines/__init__.py", line 3, in <module>
    from app.pipelines.audio_to_audio import SpeechToSpeechPipeline
  File "/app/app/pipelines/audio_to_audio.py", line 12, in <module>
    from fairseq.models.speech_to_speech.hub_interface import S2SHubInterface
ModuleNotFoundError: No module named 'fairseq.models.speech_to_speech.hub_interface'

I cannot see fairseq.models.speech_to_speech.hub_interface was never a thing in fairseq for at least most recent stable tags...

bayartsogt-ya avatar Nov 03 '22 17:11 bayartsogt-ya

I have no idea. Maybe git blame can help ?

It's possible it's an old relic of code when fairseq hub support was on a specific branch ?

Narsil avatar Nov 03 '22 17:11 Narsil

looks like changes were recent ones. https://github.com/huggingface/api-inference-community/commit/566caa5920f7f55f8b356be39a30b6737df6b636

bayartsogt-ya avatar Nov 03 '22 17:11 bayartsogt-ya

cc: @cndn I think it is added to both api-inference-community and fairseq same day but somehow fairseq changes were reverted... https://github.com/huggingface/api-inference-community/commit/566caa5920f7f55f8b356be39a30b6737df6b636 https://github.com/facebookresearch/fairseq/pull/4787

bayartsogt-ya avatar Nov 03 '22 18:11 bayartsogt-ya